Google in the Post-Jeff Dean, Post-Demis Hassabis Era

Google just reshuffled the top of its AI org chart, and the numbers behind the reshuffle are more interesting than the names. Hassabis moves from DeepMind CEO to Alphabet's first chief scientist, handing day-to-day control to Koray Kavukcuoglu. Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le are all leaving to start Discovery Loop, a public benefit corporation aimed at automating scientific research, funded partly by Google and running on Google Cloud. Wall Street didn't love it: Alphabet lost roughly $200 billion in market cap on the day.

But a former Google forecaster who spent eight years there, from 2014 to 2022, argues the market is reading this wrong in both directions. The talent exodus, he says, probably won't gut DeepMind the way headlines imply — research from an analysis of 157 frontier releases and 171 personnel moves actually found labs with more senior departures tend to build better follow-up models, likely because the real moat is the training infrastructure and the hundreds of anonymous mid-career engineers who never get memed. His own forecast puts UK headcount growing 9% by fiscal 2027, not shrinking.

What he thinks the market missed is the timeline problem. He'd already predicted Gemini 3.5 Pro for early July; Google reportedly scrapped and rebuilt the base model after coding results disappointed, and it never shipped. Now, with pretraining on Gemini 4 apparently starting in late July on Google's biggest compute budget yet, he's forecasting a median launch date of May 15, 2027 — conveniently timed for Google I/O — with 2026 launch odds falling outside his 80% confidence range entirely. For comparison, GPT-6 finished pretraining back in March and still hasn't shipped; OpenAI apparently rerouted those gains into a point release and restarted the bigger model. His read: everyone's confusing incremental updates with genuine new generations, and genuine new generations keep slipping because labs restart them when the base model disappoints.

The more unsettling thread is what happens to DeepMind's original guardrails now that it's just another SVP-led division inside Google, rather than a semi-independent lab with its own ethics board. The 2014 acquisition deal reportedly gave an ethics board veto power over any AGI DeepMind built, plus a pledge against military applications. Both have quietly eroded — Google dropped the weapons pledge from its AI principles in 2025, and in April reportedly signed a classified Pentagon deal opening Gemini up for broad government use, including human-supervised target selection. He's forecasting a 66% chance that by the end of 2027 a Gemini model gets credibly reported in an actual weapons application, which is precisely the scenario the original 2014 pledge was written to prevent.

He also thinks the more durable Google AI story isn't Gemini at all, it's cloud rent. Anthropic has reportedly committed something like $200 billion to Google Cloud compute over five years starting in 2027, dwarfing Gemini's own API and subscription revenue, which sits closer to $15 billion annualized. Discovery Loop will lean on the same infrastructure. Google may simply end up as AI's landlord rather than its frontrunner, collecting rent from the labs actually racing ahead.